IBM Granite 4.0 H Micro is a hybrid Mamba-Transformer model from IBM, and at $0.017 input per Mtok on Cloudflare Workers AI it is the single cheapest input token in the models.dev registry. Granite 4.0 H Micro is not a build model. It is the model that makes Atlas's overhead calls free. Cloudflare Workers AI and OpenRouter both list it at $0.017/$0.112 per Mtok, and Cloudflare AI Gateway at $0.017/$0.11, across a 131,000 token context window. Give it titles, commit messages, session summaries, and classification, not a multi-file refactor.
Why is IBM Granite 4.0 H Micro the cheapest input token available?
IBM Granite 4.0 H Micro costs $0.017 per Mtok input, the lowest non-zero input price anywhere in the models.dev registry, roughly half the $0.03 that DeepInfra charges for GPT OSS 20B. Cloudflare Workers AI, OpenRouter, and Cloudflare AI Gateway all quote that same $0.017 figure.
Granite 4.0 H Micro is priced the way it is because of what it is: a micro model built on a hybrid Mamba-Transformer stack. The hybrid Mamba-Transformer architecture keeps memory growth near linear with sequence length, which is why a micro model can carry a 131,000 token window at all. A pure transformer at this parameter count would choke on the KV cache long before 131,000 tokens. For Atlas, cheap input plus a large window is precisely the shape you want in the `small_model` slot, because summarization calls are input-heavy and output-light. At $0.017 in and $0.112 out per Mtok, the entire overhead category stops appearing on your invoice.
What should IBM Granite 4.0 H Micro actually do inside Atlas?
IBM Granite 4.0 H Micro should handle titles, commit messages, session summaries, and classification inside Atlas. Granite 4.0 H Micro is a micro model at $0.017/$0.112 per Mtok, and the point of the entry is that the cheap slot stops showing up on your invoice at all.
Atlas generates a steady stream of small, boring calls. Granite 4.0 H Micro is built for exactly those. Set `"small_model": "cloudflare-workers-ai/@cf/ibm-granite/granite-4.0-h-micro"` in atlas.json and leave `"model"` on a real coding model. Commit messages are a particularly good fit, because Atlas reads git branches, status, and diffs, and can stage and create commits on your behalf, so the model is summarizing a diff Atlas already computed rather than reasoning about code from scratch. Classification and routing decisions inside subagent dispatch land in the same bucket. What Granite 4.0 H Micro must not do is write the diff. It is a micro model. Give it titles, commit messages, session summaries, and classification, not a multi-file refactor.
Does IBM Granite 4.0 H Micro have a reasoning mode?
IBM Granite 4.0 H Micro has no reasoning mode. If you need chain-of-thought at this price point, GLM-4.7-Flash on Cloudflare ($0.0605/$0.40 per Mtok) is the better trade. Granite 4.0 H Micro trades thinking for the cheapest input token in the registry at $0.017 per Mtok.
The absence of a reasoning mode in Granite 4.0 H Micro is a deliberate positioning choice, not an oversight. Reasoning burns output tokens, and Granite 4.0 H Micro's whole pitch is that its tokens cost nearly nothing. If the work in your `small_model` slot genuinely needs a thinking trace, paying $0.0605/$0.40 per Mtok for GLM-4.7-Flash on Cloudflare is the honest upgrade, at roughly 3.5x the input cost. If the work is summarization and classification, which is what the slot usually holds, Granite 4.0 H Micro at $0.017/$0.112 per Mtok is strictly cheaper for the same result. Atlas lets you switch the active model and provider on the fly with favorites and recents, so testing both against your own workload is a small experiment.
Why do enterprises pick IBM Granite over a cheaper open model?
IBM ships Granite under Apache 2.0 with a documented training-data provenance trail, which matters when legal has an opinion about what trained your coding assistant. Granite 4.0 H Micro at $0.017/$0.112 per Mtok is cheap, but provenance, not price, is why IBM is on the shortlist.
Granite 4.0 H Micro's Apache 2.0 license and IBM's published data-provenance documentation are load-bearing for regulated teams. A coding assistant reads proprietary source code, and the question of what data trained the model that read it is one some organizations must answer on paper. Atlas gives you a second lever on the same concern: Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so indexing and inference can be separated. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, which gives the same audience an enforcement story to go with the provenance story.
How do you configure IBM Granite 4.0 H Micro in Atlas?
Configuring IBM Granite 4.0 H Micro in Atlas takes three steps: export CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_KEY, run `atlas models cloudflare-workers-ai`, and set the `small_model` key. The model resolves as `@cf/ibm-granite/granite-4.0-h-micro` at $0.017/$0.112 per Mtok.
Granite 4.0 H Micro is served through Cloudflare Workers AI, so authentication uses both CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_KEY rather than a single key. Run `atlas models cloudflare-workers-ai` and find the `@cf/ibm-granite/granite-4.0-h-micro` row before editing atlas.json, so a typo in the model id surfaces as a missing row rather than a runtime failure. Then set `"small_model": "cloudflare-workers-ai/@cf/ibm-granite/granite-4.0-h-micro"`. Leave `"model"` on a real coding model. Cloudflare AI Gateway carries the same model at $0.017/$0.11 per Mtok if you are already routing through the gateway.
Setup
- 01Export CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_KEY.
- 02Run `atlas models cloudflare-workers-ai` and find the `@cf/ibm-granite/granite-4.0-h-micro` row.
- 03Set `"small_model": "cloudflare-workers-ai/@cf/ibm-granite/granite-4.0-h-micro"` in atlas.json.
- 04Leave `"model"` on a real coding model. The point of this entry is that the cheap slot stops showing up on your invoice at all.
- 05If your `small_model` work needs chain-of-thought, compare GLM-4.7-Flash on Cloudflare at $0.0605/$0.40 per Mtok, since Granite 4.0 H Micro has no reasoning mode.
Frequently asked questions
- how much does ibm granite 4.0 h micro cost
- IBM Granite 4.0 H Micro is $0.017/$0.112 per Mtok on Cloudflare Workers AI and OpenRouter, and $0.017/$0.11 per Mtok on Cloudflare AI Gateway. The $0.017 input price is the lowest non-zero input price in the registry.
- what is the cheapest llm input price in 2026
- IBM Granite 4.0 H Micro at $0.017 per Mtok input is the lowest non-zero input price anywhere in the models.dev registry, roughly half the $0.03 that DeepInfra charges for GPT OSS 20B.
- what is the context window of granite 4.0 h micro
- Granite 4.0 H Micro holds 131,000 tokens on both Cloudflare Workers AI and OpenRouter. The hybrid Mamba-Transformer architecture keeps memory growth near linear with sequence length, which is why a micro model can carry that window.
- can ibm granite 4.0 h micro write code in a coding agent
- Not well. Granite 4.0 H Micro is a micro model. Give it titles, commit messages, session summaries, and classification, not a multi-file refactor. Leave a real coding model in the `"model"` slot.
- does granite 4.0 h micro support reasoning
- No. Granite 4.0 H Micro has no reasoning mode. If you need chain-of-thought at this price point, GLM-4.7-Flash on Cloudflare at $0.0605/$0.40 per Mtok is the better trade.
- how do I use granite 4.0 h micro with atlas
- Export CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_KEY, run `atlas models cloudflare-workers-ai` to find the `@cf/ibm-granite/granite-4.0-h-micro` row, then set `"small_model": "cloudflare-workers-ai/@cf/ibm-granite/granite-4.0-h-micro"` in atlas.json.
- is ibm granite open source and apache licensed
- Yes. IBM ships Granite under Apache 2.0 with a documented training-data provenance trail, which matters when legal has an opinion about what trained your coding assistant.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Gemini CLI: A Developer's Guide to Terminal AI Agents in 2026
Atlas vs Gemini CLI in 2026: Compare terminal AI coding agents. Atlas offers permission-gated tool calls and diff-based approvals. Gemini CLI provides a million-plus token context and a free tier.
Atlas for Julia: A Terminal-Native AI Coding Agent for Project.toml Packages in 2026
Atlas is a terminal-native AI coding agent for Julia in 2026. It reads dispatch signatures and Project.toml [deps], fixes type instabilities, runs Pkg.test(), and applies JuliaFormatter.
Atlas for COBOL: Copybooks, PIC Clauses, and GnuCOBOL in 2026
Atlas is a terminal-native AI coding agent for COBOL in 2026. It reads your divisions and copybooks, explains a paragraph before touching it, and compiles with cobc under GnuCOBOL.
Atlas for Bun: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Bun. Swap Node compatibility shims for Bun.serve and bun:sqlite, iterate with bun test, and format with bun fmt in 2026.
Atlas vs Claude Code: Terminal AI Coding Agents in 2026
Compare Atlas and Claude Code in 2026 for terminal AI coding. Atlas offers free core, BYO model keys, and robust diffing, while Claude Code provides deep Claude tuning and strong multi-step execution.
Atlas for Deno: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Deno. Tighten --allow-all down to real permission scopes, add JSR deps, and run deno test and deno check in 2026.
Atlas for PHP in 2026
Atlas, the terminal-native AI coding agent, empowers PHP developers in 2026 with intelligent code understanding, secure workflows, and direct integration for Composer and PSR standards.
Atlas for Axum in 2026
Atlas is a terminal-native AI coding agent for Axum in 2026. It decodes tower trait-bound errors, adds IntoResponse types, and runs cargo nextest run.